A few days ago I wrote about CosmoBot, my personal AI assistant built on Clawdbot. Morning briefs via Telegram. Check-ins throughout the day. Everything stored in my Obsidian vault. It’s been… transformative? That word gets overused, but it fits.

The thing is, I don’t really understand how it works.

Clawdbot—which became Moltbot, which became OpenClaw; two name changes in a week—is a big codebase. I checked out the repo with every intention of contributing. I poked around, and quickly felt overwhelmed. There’s a lot going on in there, and the number of contributions every day is mind-boggling. The age of AI-powered engineering is here, and we’re watching it take shape in realtime.

I could have just kept using OpenClaw—and I still might. But I can’t use something I don’t understand. Even as a kid, I’d take apart new toys to see how they worked. That habit never went away.

So I’m building my own agent from scratch.

Well, not scratch scratch. I’m using the Anthropic Agent SDK. But I’m building the whole thing myself, using OpenClaw’s code as a reference to learn from.

His name is James. After 007.

Here’s what surprised me: I had James up and running in less than 30 minutes.

The core system is simple. Like, really simple. The SDK handles the plumbing. The true magic comes from the LLM itself. You’re not building intelligence—you’re just giving it hands.

And once James was online, he started reviewing his own code and suggesting improvements. The agent helping build the agent. Recursive in the best way.

—

Now I’m building a web client to make the UX nicer. Adding tools so James can access my calendar and other apps. I’ve even got a fairly robust safety layer built in—you don’t give an AI agent access to your life without guardrails.

CosmoBot showed me what’s possible. James is how I’m learning to build it myself.